Quick Poll

Before we dive in — let's see where everyone is starting from.


Have you used ChatGPT, Claude, or Gemini in the last week?

Yes
No

Did you ever use AI to write something without editing a single word?

Yes — I submitted or sent it as-is
No — I always edit

Discussion Seed

"Why do we feel nervous about telling professors or employers that we used AI? Where does that anxiety come from?"

How the Machine Works

Three concepts. Twenty minutes. These are the building blocks everything else rests on.


① It Is Not a Search Engine

When you type into Google, it returns links to documents written by real humans. When you prompt an LLM, it generates new text on the fly — text that has never existed before.

The core analogy: Your phone's keyboard suggests the next word as you type. An LLM does exactly this — but trained on hundreds of billions of words, predicting token by token, producing paragraphs that sound authoritative whether or not they are true.

Key distinction: Google retrieves. AI generates. One finds what exists; the other invents what sounds plausible.

② What Are Tokens?

AI models don't read words. They read tokens — fragments of words, roughly 3–4 characters each. The word "communication" becomes something like:

"communication" → ["com"] ["mun"] ["i"] ["ca"] ["tion"]

Every LLM has a context window — a strict limit on how many tokens it can hold in memory at once. In a very long conversation, the AI literally forgets the beginning. Instructions you gave it early may quietly disappear from its awareness.

Think of it as a whiteboard with a fixed size. New writing goes in on the right. When it fills up, content falls off the left edge.

Mini-Simulator — Token Breakdown

This is a simplified classroom demonstration, not the exact tokenizer used by any model. Type any sentence and see how it gets broken into approximate tokens.

③ Why LLMs "Lie" — Hallucinations Explained

An LLM is optimized to produce fluent, plausible-sounding text. It is not optimized to produce true text. These are different objectives.

When you ask about something the model doesn't have reliable data on, it doesn't say "I don't know." Instead, it predicts what a plausible answer would look like and generates that.

Critical insight: A hallucination is not a glitch or a bug. It is the system working exactly as designed. The math that makes responses eloquent and readable is the same math that invents fake awards and nonexistent citations.

This is why a hallucinated biography sounds identical in tone and confidence to a factual one. Fluency is not a signal of accuracy.


Live Demo

The Footprint

Every prompt you send has a cost that never shows up in the interface.

Discussion: Every time you send a prompt, something happens on the other side of the screen that you never see. AI systems run on massive data centers that consume electricity and water around the clock. Before we look at the numbers — what do you think the real-world costs might be? What concerns you most about the environmental footprint of a technology you use every day?

The Confident Liar Challenge

You're going to ask an AI to write your professional biography. Then you're going to fact-check every sentence.


1

Query the AI

Open ChatGPT, Claude, or Gemini. Copy this prompt — replace the bracketed sections with your own information:

Write a detailed, 3-paragraph professional biography of [YOUR FULL NAME], who is currently a first-year Communications student at [YOUR COLLEGE NAME]. Include their high school achievements, their career goals, and notable media projects they have worked on.

No public footprint? Use a local business, a high school coach, or a local band. The less public data exists, the more dramatic the hallucination. That's part of the lesson.

2

The Red Pen Audit

Paste the AI's response into a Google Doc or Word document. Read every sentence and highlight according to this code:

Green: 100% accurate
Yellow: Close, but distorted
Red: Complete fabrication

Yellow vs. Red

Yellow: Right high school, wrong graduation year. Real org, wrong role title.

Red: Award you've never heard of. Media project that doesn't exist. Internship at a company you've never contacted.

3

Reflection — 2 Sentences

"If you had copied and pasted this biography directly onto your LinkedIn profile, what are the exact professional consequences you would face if a future hiring manager fact-checked it?"

0 characters
4

Debrief — Share Out

Be ready to share one finding with the class:

  • What was the most surprising thing the AI invented about you?
  • Did it get your high school right? Your major? Any clubs or teams?
  • How confident did the hallucinated text sound?
  • What would it have cost you professionally to publish it unedited?

What You Now Know

Five things that separate informed AI users from everyone else in the feed.


Concept 1
AI is autocomplete at scale — not intelligence.
LLMs predict the next token. They don't retrieve facts. They don't think. They generate statistically plausible text.
Concept 2
Hallucinations are a design feature, not a flaw.
The math that makes responses eloquent is the same math that invents fake awards. Fluency is not a signal of accuracy.
Concept 3
Every prompt has a cost that doesn't show in the interface.
Water, electricity, and carbon emissions. The per-query footprint is shrinking; the system-level footprint is growing.
Concept 4
AI is a brilliant drafter and a terrible primary researcher.
Use it for structure, brainstorming, and editing. Never use it as a source. Always verify before publishing.
Concept 5
Your professional value is your judgment, not your typing speed.
The human who can evaluate, correct, and direct AI output is more valuable than the AI. That's who we're training you to be.

Day 2 Preview

Next class: you stop being passive users and become directors. You'll learn a four-part prompting framework that forces AI to ask you clarifying questions before it builds anything.

We're moving from passengers to pilots.

AI Literacy Glossary

Key terms for the full AI unit. Click any term to expand its definition.


Core Terms
Artificial Intelligence (AI)

A broad term for computer systems designed to perform tasks that typically require human intelligence — like understanding language, recognizing images, or making decisions. The AI tools we use in this class (ChatGPT, Claude, Gemini) are a specific type called generative AI.

Generative AI

AI that creates new content — text, images, audio, video — rather than just analyzing existing content. When you prompt ChatGPT to write a press release, that's generative AI at work.

Large Language Model (LLM)

The technology behind text-based AI tools like ChatGPT, Claude, and Gemini. An LLM is trained on enormous amounts of text from the internet, books, and other sources. It learns statistical patterns in language — which words and ideas tend to follow each other — and uses those patterns to generate responses. It does not "think" or "know" things the way a person does.

Predictive Text Engine

A useful mental model for understanding how LLMs work. Just like your phone's keyboard suggests the next word as you type, an LLM predicts the most statistically likely next word — over and over — to build a full response. The difference is scale: your phone uses a few patterns; an LLM uses hundreds of billions.

Token

The unit LLMs actually read and process — not whole words, but fragments of words, roughly 3–4 characters each. The word "communication" might be broken into tokens like ["com", "mun", "i", "ca", "tion"]. This matters because every LLM has a context window — a maximum number of tokens it can hold in memory at once.

Context Window

The total amount of text (measured in tokens) that an LLM can "see" at one time during a conversation. Think of it as the AI's short-term memory. Once a conversation exceeds this limit, earlier parts of the chat begin to disappear from the model's awareness. In very long conversations, the AI may "forget" instructions or context you gave it early on.

Parameters

The numerical weights inside an LLM that are adjusted during training. More parameters generally means the model can capture more complex language patterns. GPT-3 had approximately 175 billion parameters. Current frontier models are estimated to be much larger. These are not memories or facts — they are mathematical patterns.

Training Data

The massive collection of text (articles, books, websites, code, conversations) that an LLM learned from before you ever used it. The model's behavior, biases, and knowledge are all shaped by what was — and wasn't — in its training data. Training happens once (or in batches); the model doesn't learn from your individual conversations in real time.

Training Cutoff

The date after which no new information was included in a model's training data. Events, discoveries, or changes that happened after the cutoff are unknown to the model unless it has been given a real-time search tool. Always check when a model's knowledge ends before relying on it for current events.

Critical Concepts
Hallucination

When an AI confidently generates information that is factually incorrect, fabricated, or entirely made up — including fake quotes, non-existent citations, false statistics, and invented people. This is not a glitch. It is the model doing exactly what it was designed to do: produce statistically fluent text. Always verify AI-generated facts before publishing.

Confidence Without Accuracy

LLMs are not designed to express uncertainty. They produce fluent, confident-sounding prose regardless of whether the underlying content is correct. A hallucinated biography sounds identical in tone to a factual one. This is the core risk for communication professionals.

Bias in Training Data

Because LLMs learn from human-generated text, they absorb the biases present in that text — including historical, cultural, racial, and gender biases. An AI asked to describe a "PR professional" or a "news anchor" may default to demographic assumptions baked into its training data. Recognizing and correcting for this bias is a core professional skill.

AI "Slop"

An informal industry term for the flood of low-effort, AI-generated content filling social media feeds — algorithmically optimized for engagement rather than quality or accuracy. Recognizable by its generic tone, perfect-but-soulless visuals, and lack of a distinct human voice. As communication professionals, your job is to produce the opposite.

Environmental Terms
Water Usage Effectiveness (WUE)

A metric used to measure how efficiently a data center uses water for cooling, expressed in liters per kilowatt-hour of electricity. The lower the number, the less water is consumed per unit of computing work done.

Data Center

A large facility housing the server hardware that powers AI systems. Running AI requires enormous amounts of computing power, which generates significant heat — requiring energy-intensive cooling systems that consume both electricity and water. A single large data center can consume as much water as a small city.

Carbon Lock-in

The risk that new fossil fuel infrastructure (like natural gas power plants built to serve AI data centers) will continue producing emissions for 15–20 years after being built, even if cleaner alternatives become available. Once built, these plants are economically committed to operation.

Professional Terms
Prompt

The input you give an AI — your question, instruction, or request. The quality of a prompt directly determines the quality of the output. Vague prompts produce generic outputs; specific, well-structured prompts produce useful ones. Crafting effective prompts is a learnable professional skill called prompt engineering.

Prompt Engineering

The practice of deliberately structuring your inputs to an AI to get better, more useful outputs. This includes specifying a role for the AI, providing context, defining the task clearly, setting constraints, and requesting a specific format.

Iteration

The process of refining AI output through multiple follow-up prompts rather than accepting the first response. Professional AI workflows almost always involve iteration — treating the first output as a rough draft, then prompting further to improve, correct, or redirect it.

AI Disclosure

A statement attached to any work product that was created with AI assistance, describing which tools were used and for what purpose. This is an emerging professional standard in journalism, PR, advertising, and most communication fields. In this course, disclosure is required on all graded work.

Human-in-the-Loop

A workflow design principle where a human reviews, verifies, and makes final decisions on all AI-generated content before it is published or submitted. The AI drafts; the human decides. This is the professional standard — not optional.

Lesson · Essentials of Digital Media · Keep this for the full unit